System
Code repairs itself while you sleep.
Triages incidents, writes patches, opens PRs. Ed25519-signed audit, 10/10 threat model mitigated, evidence pack for SOC2 / ISO 27001 / NIS2 / GDPR generated in one click. Reliability AI for teams that can't compromise on code confidentiality.
I
Reliability as a measure, not a story
Hermes defines reliability as the speed at which the system reacts to an error — not the absence of errors. The mark itself says it: the partial derivative of H over time is the KPI you take to the board. Every other number is a consequence of that one.
II
Seven stages, only one touches a human
The engine follows a seven-stage pipeline: StackDetector, Planner, Applier, BaselineCollector, Triager, Patcher, Validator. Only the last — when policy blocks an automatic merge — calls a human review via step-up Face ID. Everything else runs twenty-four hours a day, sentinel active, median latency under one minute.
III
10/10 threat model mitigated
The security audit covers ten vectors: prompt injection, branch escape, PR XSS, secret leak, webhook spoofing, rate-limit bypass, auth replay, trigger fingerprint replay, audit log tampering, model output exfiltration. Each mitigation is tracked to release version, source file, test coverage. Signed PDF available on request.
IV
Append-only audit, evidence pack in one click
The audit log runs on PostgreSQL with REVOKE UPDATE, DELETE FROM PUBLIC — INSERT and SELECT only. Every event Ed25519-signed. The CISO clicks "Generate evidence pack" and gets a dossier compliant with SOC2 Type II, ISO 27001 A.12.6, NIS2 art.21 (Italian D.Lgs 138/2024), and GDPR art.32. One engine, four frameworks, zero opacity.
V
Keyless by design
Hermes never asks for a personal-access-token. Each repository has its own Ed25519 deploy key, public on GitHub, private encrypted AES-256-GCM at-rest. Leak surface is zero by design. Per-repo revocation, scheduled rotation, no broad OAuth scopes on the organization.
Capabilities
01
Deduplicated trigger fingerprint
Every incident has a stable signature: SHA-256 fingerprint of relevant symptoms. No duplicated work, no infinite patching on the same intermittent error.
02
AI triager with trust score
Classifies every trigger in seconds: noise, actionable, transient. Patch decision above 0.55 trust threshold, human escalation below. Reasoning hashed and signed.
03
Generative Patcher · max 3 attempts
The configured model (sonnet-4.5 by default, haiku-4.5 for cost-tier) proposes up to three patches. Each attempt audited. Four = mandatory escalation.
04
Validator sandbox
Runs test suites in ephemeral container, measures coverage delta and regression, blocks merges that violate the no-regression rule. Verdict in seconds.
05
YAML policy engine
Readable, versioned rules: cost cap, sanitized branch, no-regression, secret leak prevention. Every change is one signed line in the log.
06
Face ID step-up review
When policy blocks automatic merge, the operator gets a critical push and does WebAuthn biometric step-up. Sixty-second TTL, audit signed on identity.
07
Native multi-tenant
Tenant slug, per-tenant policy, per-tenant billing, isolated audit. System integrators can orchestrate multiple end customers from a single console.
08
Web console and mobile PWA
Six essential views on the web (Workflows, Audit, Baseline, Threat, Discovery, Settings). Critical push on iPhone for escalated reviews at three in the morning.
Technology stack
TypeScript engine, over six hundred green tests, ephemeral Docker container for every sandbox, append-only Postgres for audit, WebAuthn for human authorization. LLM models chosen per task — no lock-in.
- hermes-core (TypeScript)Pipeline orchestration · 7 stages
- Next.js 16 · App RouterMulti-tenant web console
- PostgreSQL 16 · append-onlyAudit log with REVOKE UPDATE/DELETE
- Ed25519Cryptographic signature on every event
- AES-256-GCM · KMS rotatingDeploy key encryption at-rest
- WebAuthn passkeyAuth + step-up review · Face ID/Touch ID
- Claude Sonnet 4.5 · Haiku 4.5LLM models · Triager + Patcher
- Docker · ephemeral sandboxIsolated Validator runtime
- GitHub Actions / Sentry / OTelStandard trigger sources
- policy_engine.yamlVersioned per-tenant rules · Git-ops
- SSE · 20s keepaliveLive workflow stream in Console
- @oasi/core-aiLibrary shared with Atena and Avatar
Delivery roadmap
Four independent phases. The customer sees value before approving the next. The Sentinel phase only goes live after Hermes has learned the tenant's patterns — no shortcuts.
01 · 1 week
Discovery
StackDetector reads the repository — no LLM, no writes. Output: StackReport with detected frameworks, present telemetry, test coverage, warnings. Signed Statement of Work.
02 · 2-3 weeks
Telemetry onboarding
Planner plans onboarding actions (Sentry, OpenTelemetry, CI hooks). Applier opens the onboarding PR on the repository. From here on, Hermes sees the incidents.
03 · 4-8 weeks
Baseline
BaselineCollector observes the system's noise. Cluster classification: noise vs actionable vs transient. Recommended thresholds computed on the tenant's real data.
04 · Continuous
Active Sentinel
Triager + Patcher + Validator + PolicyEngine in production twenty-four hours a day. Signed audit, quarterly evidence pack, automatic Quarterly Board Report PDF generation.